automated extraction of psychological symptoms from French ORFeAD forensic medical certificates for interpersonal violence victims

Source article: Rule-Based Versus Generative Extraction of Psychological Symptoms From Forensic Medical Certificates: A Feasibility Study in the French ORFeAD Network

Forensic medical certificates for victims of interpersonal violence describe psychological symptoms in unstructured prose, rarely explicit and requiring inference. To assess the feasibility of two extraction pipelines-rule-based and a locally served generative large language model-in a French multicentric forensic network, and to identify variables usable without human verification. 110 randomly selected 2021 certificates from 12 ORFeAD units were processed by both pipelines: rule-based (segmentation, lexicons,…

By the Truvace deskOctober 5, 20262 min read1 sourcePeer-reviewedTRV-2026-1282· cite & verify

Rule-Based Versus Generative Extraction of Psychological Symptoms From Forensic Medical Certificates: A Feasibility Study in the French ORFeAD Network — Granada - Hospital Doctor Olóriz 3 by Zarateman. CC0 · http://creativecommons.org/publicdomain/zero/1.0/deed.en

Trace impact readingNegative state

P 71

Both sides are scored from claims and sources, not community votes.

G 66

In brief

Researchers tested two automated pipelines on 110 randomly selected 2021 forensic medical certificates from 12 units of the French ORFeAD network, comparing a rule-based system using segmentation, lexicons and negation detection to a locally served Llama 3 8B model via Ollama, against physician coding of 35 binary variables.

The comparison matters because forensic certificates contain psychological symptoms in unstructured prose that requires inference, and reliable extraction could enable multicentric research on victim outcomes; however the feasibility results show uneven sensitivity, higher energy and GPU requirements for the generative approach, and explicit limits against using either pipeline for individual-level decisions without per-variable qualification.

Main points

110 randomly selected 2021 certificates from 12 ORFeAD units were processed by both pipelines and coded by two forensic physicians with third-party arbitration.
Rule-based pipeline used segmentation, lexicons, regular expressions, negation detection; generative pipeline used Llama 3 8B, 4-bit, Ollama, near-default settings.
35 binary variables, 23 psychological or subjective, were evaluated for sensitivity, specificity and accuracy with 85% reliability defined as usable.

The gain

Automated extraction of psychological and subjective variables from unstructured forensic certificates was feasible with high specificity, with 24 of 35 variables meeting an 85% reliability threshold under at least one pipeline.

The problem

Sensitivity was heterogeneous and low for low-prevalence symptoms, 11 of 35 variables failed the reliability threshold under either pipeline, and neither pipeline supports individual-level decisions, with the generative model incurring far higher compute cost.

The rundown

The study compared a lexicon and regex rule-based system requiring 0.1-0.15 CPU-seconds per document against a locally served Llama 3 8B 4-bit model requiring 20-60 GPU-seconds, finding the local model did not clearly outperform at three orders of magnitude greater energy cost.

Authors coded 35 binary variables including 23 psychological or subjective ones described in unstructured prose that is rarely explicit and requiring inference, and concluded per-variable qualification is needed before any research use.

What this doesn’t fix

Feasibility design with 110 certificates from 2021, no inferential testing, and out-of-the-box LLM settings limits generalizability and requires per-variable qualification before research use.